Tiangong Omni's face-covering sprint reportedly emerged from reinforcement learning in NVIDIA's Isaac stack

A shy-looking robot takes first place

A clip of the Tiangong Omni humanoid robot sprinting with both hands pressed over its face — and finishing first — circulated widely after researcher David Ha shared it on X roughly a week earlier, according to a write-up on dev.to. The run took place at a robot games event in Beijing, and the oddly bashful posture became the immediate talking point. The technically interesting part, though, is where the gait came from.

Not designed, but discovered

The dev.to write-up reports that engineers at the Beijing Humanoid Robot Innovation Center, known as X-Humanoid and the team behind the Tiangong robot line, said the pose was not a pre-designed animation. Instead, it emerged on its own inside the simulation environment where the robot trains — the "metaverse" the company uses as its training ground.

As the write-up describes it, the pipeline runs through NVIDIA's robotics stack:

• Omniverse serves as the underlying 3D and digital-twin platform.
• Isaac Sim runs on top of it as the physics simulator, where millions of parallel reinforcement-learning trial-and-error episodes play out.
• Isaac Lab, NVIDIA's open-source reinforcement-learning framework, handles the large-scale parallel training.
• Tien-Kung-Lab, an open-source framework X-Humanoid built on top of Isaac Lab, provides the locomotion-control algorithms for the Tiangong series.

Once a policy is trained, it is exported and cross-checked in MuJoCo — a Sim2Sim validation step — before being deployed to the physical Tiangong Omni, with a small amount of fine-tuning on the real machine.

A familiar phenomenon, now on real legs

David Ha, quoted in the write-up, compared it to "novel locomotion policies discovered in MuJoCo environments, except that they work in the real world." Reinforcement-learning researchers have long watched agents invent strange but effective ways of moving inside simulators. The unusual part here is that such a policy survived the transfer to physical hardware and won a public race.

Two simulators, one pipeline

The write-up also places the result in the broader tooling landscape. The Tiangong team's main training runs in the NVIDIA ecosystem, with thousands of parallel environments on GPU. The same robot model and reward functions are then moved into MuJoCo to reproduce the gait — a deliberate cross-check for whether the PhysX-based simulation had introduced artifacts, and a way to probe the sim-to-real gap before touching the real robot.

The two ecosystems also interoperate more than they used to, according to the write-up: Isaac Sim can import MuJoCo's MJCF model format directly, and Isaac Lab 3.0 can run MuJoCo-Warp as a physics backend without launching Omniverse rendering. The face-covering gait could in principle be trained in either environment; the Isaac stack was chosen for the heavy lifting because of its large-scale parallelism and sensor simulation.

One caveat worth stating plainly: these training details come from a single write-up relaying the company's engineering claims alongside Ha's post, so they should be read as reported rather than independently verified.

Why it matters

The headline here is emergent behavior that survived sim-to-real transfer. Reinforcement learning routinely produces odd-but-functional gaits in simulation, and most of them collapse once gravity, friction and hardware tolerances get involved. A policy that not only transfers to a physical humanoid but wins a race is evidence that massive parallel training, followed by cross-simulator validation, is becoming a dependable route to locomotion that no one has to hand-design.

It also illustrates an industry pattern that is hardening into standard practice: train where parallelism and sensor simulation are cheap, validate where contact dynamics are trusted. As more humanoids are developed this way, distinguishing robust real-world capability from simulator-flavored accidents will matter increasingly — which is precisely the job of the MuJoCo cross-check in X-Humanoid's workflow.